Severity Classification of Mental Health Related Tweets

The use of social media has drastically gone up over the last decade. With this comes more opportunity and also more problems. There is a rise in the number of mental health-related issues, and it is to some extent possible to detect such cases via user posts and tweets (in our case). Previous research has focused on classifying mental health diseases such as depression, bipolar disorder, schizophrenia, etc., from already filtered data. However, not much has been done to filter out tweets that might be sarcastic, which are generally misclassified, or tweets that might not be intended in a harmful way and, in general, classify tweets based on their severity with regards to mental health. This paper uses multiple models to classify tweets based on their severity and classify them into three classes that help determine whether they help people with mental conditions or sarcasm. We use famous neural network architectures such as Bidirectional LSTMs, GRUs, and a custom HYBRID model to carry out the classification. The models could detect sarcasm in tweets and identify tweets that were helpful despite having words like 'depression' and 'anxiety.' We obtained F1 scores of 74% on completely unseen data, which is a good starting point considering the limited available data. This paper should serve as a utility for future research in this area and act as a primary data collection and segregation filter.